Risk rechecking method and device and storage medium

Abstract

The invention relates to a risk rechecking method and device and a storage medium. The method comprises the steps that feature extraction is carried out based on at least one document sample, multiple types of feature vectors are obtained, the multiple types of feature vectors comprise at least two of text feature vectors, numerical feature vectors and graph feature vectors, and the at least one document sample does not comprise known misinformation samples or comprises part of the known misinformation samples; determining a risk review model based on the multiple types of feature vectors corresponding to each document sample; and processing the to-be-discriminated document based on the risk review model to obtain a risk review result, the risk review result being used for indicating whether the to-be-discriminated document has a false alarm. According to the embodiment of the disclosure, a large amount of structured and non-structured information contained in the receipt data can be fully utilized in a scene lacking the labeled sample, and context semantics, statistical laws and entity association relationships contained in the express receipt information can be comprehensively and deeply understood, so that the false alarm rate is effectively reduced.

Description

Risk review methods, devices and storage media Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and in particular to a risk review method, apparatus and storage medium. Background Technology

[0002] In existing customs express clearance operations, the identification of prohibited and restricted items mainly relies on target detection algorithms based on intelligent image analysis. However, due to factors such as non-standard declaration data, high complexity of commodity forms, and insufficient algorithm generalization ability, these algorithms are prone to misreporting non-prohibited and restricted items as prohibited or restricted items. Such misreporting issues lead to an increase in the workload of manual verification by customs and a extension of clearance time. Therefore, how to effectively reduce the misreporting rate has become an urgent problem to be solved in the current customs supervision field.

[0003] Currently, research on risk verification of prohibited and restricted items mainly focuses on two aspects: hardware optimization, such as increasing the radiation intensity and resolution of equipment, adding multi-angle scanning, and using dynamic rotation imaging technology; and algorithmic improvement, such as increasing the diversity of training data, capturing more dimensional image features, and adopting more advanced algorithm network structures. However, the above optimization methods have certain limitations: firstly, they rely too heavily on image data and fail to fully utilize the large amount of structured and unstructured information contained in the document data, which may be related to false alarm behavior; secondly, in practical application scenarios, the number of false alarm samples is often limited and difficult to collect, which can lead to overfitting or underfitting problems in the model. Therefore, in scenarios lacking labeled samples, how to fully mine and utilize document information to construct an auxiliary false alarm risk analysis mechanism has become an important research direction for risk verification methods of prohibited and restricted items. Summary of the Invention

[0004] In view of this, this disclosure provides a risk review method, apparatus and storage medium.

[0005] According to one aspect of this disclosure, a risk review method is provided. The method includes:

[0006] Feature extraction is performed based on at least one document sample to obtain multiple types of feature vectors. The multiple types of feature vectors include at least two of the following: text feature vectors, numerical feature vectors, and graph feature vectors. At least one document sample does not include known false alarm samples or includes some known false alarm samples.

[0007] Based on the feature vectors of multiple types corresponding to each document sample, the risk review model is determined.

[0008] The risk review model is used to process the documents to be judged and obtain the risk review results. The risk review results are used to indicate whether there are false alarms in the documents to be judged.

[0009] In one possible implementation, feature extraction is performed based on at least one document sample to obtain multiple types of feature vectors, including:

[0010] Preprocess at least one text field in a document sample to obtain a preprocessed document sample.

[0011] Text enhancement processing is performed on the target document sample in the preprocessed document sample to obtain positive samples, and some or all of the samples are selected from other document samples as negative samples.

[0012] The target document sample, positive sample, and negative sample are encoded using a text encoder to obtain the original semantic vectors corresponding to the target document sample, positive sample, and negative sample, respectively.

[0013] The parameters of the text encoder are optimized and trained based on the original semantic vectors corresponding to the target document samples, positive samples, and negative samples, and the text feature vectors corresponding to each document sample are obtained based on the trained text encoder.

[0014] In one possible implementation, the parameters of the text encoder are optimized and trained based on the original semantic vectors corresponding to the target document samples, positive samples, and negative samples, respectively. The trained text encoder then yields the text feature vectors corresponding to each document sample, including:

[0015] The first loss function value is calculated based on the original semantic vectors corresponding to the target document sample, positive sample and negative sample respectively. The parameters of the text encoder are optimized and trained based on the first loss function value. The first semantic vector corresponding to each document sample is obtained using the trained text encoder.

[0016] The first semantic vector is processed by a multilayer perceptron encoder to obtain a first latent semantic vector. The dimension of the first latent semantic vector is smaller than that of the first semantic vector.

[0017] The first latent semantic vector is processed by a multilayer perceptron decoder to obtain a second semantic vector, which has the same dimension as the first semantic vector.

[0018] The second loss function value is calculated based on the first and second semantic vectors. The parameters of the multilayer perceptron encoder are optimized and trained based on the second loss function value. The second latent semantic vector corresponding to each document sample is obtained by using the trained multilayer perceptron encoder as the text feature vector.

[0019] In one possible implementation, feature extraction is performed based on at least one document sample to obtain multiple types of feature vectors, including:

[0020] Numerical fields in at least one document sample are preprocessed according to different types to obtain multiple types of numerical features;

[0021] For any type of numerical feature, it is grouped according to a preset grouping method to obtain multiple sets of numerical features corresponding to that type.

[0022] For any type of numerical feature, based on the proportion of false alarm samples and non-false alarm samples in the corresponding document samples of each group of numerical features, at least one type of numerical feature that meets the preset conditions is selected from multiple types as a numerical feature vector.

[0023] In one possible implementation, feature extraction is performed based on at least one document sample to obtain multiple types of feature vectors, including:

[0024] Entities and relationships are extracted from at least one document sample to determine the attribute values ​​of each entity and the relationships between entities, so as to obtain a knowledge graph. Nodes in the knowledge graph are used to represent the corresponding entities, and edges in the knowledge graph are used to represent the relationships between entities. Each node is associated with a category label, which is used to indicate the category of the entity corresponding to the node.

[0025] For any target node in the knowledge graph, a graph neural network is used to process the feature vector of the target node and the feature vectors of the target node's neighboring nodes to obtain the first multidimensional feature vector.

[0026] The predicted category of the entity corresponding to the target node is determined based on the first multidimensional feature vector, and the third loss function value is calculated based on the predicted category and the category label of the target node.

[0027] The parameters of the graph neural network are optimized and trained based on the value of the third loss function, and the second multidimensional feature vector corresponding to the knowledge graph is obtained by using the trained graph neural network.

[0028] For any document sample, the feature vectors of the nodes corresponding to any document sample in the second multidimensional feature vector are combined to form the graph feature vector corresponding to that document sample.

[0029] In one possible implementation, a risk review model is determined based on feature vectors of multiple types corresponding to each document sample, including:

[0030] In the case where at least one document sample includes a known false alarm sample, at least one set of sample pairs and the label value corresponding to each sample pair are determined based on at least one document sample. The types of sample pairs include: a first sample pair consisting of two known false alarm samples, a second sample pair consisting of a known false alarm sample and an unknown sample, and a third sample pair consisting of two unknown samples. The label value of the first sample pair is greater than the label value of the second sample pair, and the label value of the second sample pair is greater than the label value of the third sample pair.

[0031] For any pair of samples, input the feature vectors corresponding to each document sample in the pair into the end-to-end learner to obtain the output vector;

[0032] The output vector is input into the ordinal regressor to obtain the predicted value;

[0033] The fourth loss function value is calculated based on the label value and the predicted value. The parameters of the end-to-end learner and the ordinal regressor are optimized and trained based on the fourth loss function value to obtain the risk review model.

[0034] In one possible implementation, the document to be judged is processed based on a risk review model to obtain the risk review result, including:

[0035] From at least one document sample, some known false alarm samples are extracted and combined with the document to be judged to form multiple first sample pairs to be judged; and from at least one document sample, some unknown samples are extracted and combined with the document to be judged to form multiple second sample pairs to be judged.

[0036] Based on the risk review model, multiple first and second sample pairs to be discriminated are processed to obtain the corresponding first and second discriminant scores.

[0037] The risk review results are determined based on the first and second discriminant scores.

[0038] In one possible implementation, a risk review model is determined based on feature vectors of multiple types corresponding to each document sample, including:

[0039] In the case that at least one document sample does not include known false alarm samples, an isolation tree is constructed based on the feature vectors of each type corresponding to each document sample, wherein each leaf node in the isolation tree corresponds to the feature vector of a document sample.

[0040] Use at least one isolated tree as the risk review model;

[0041] The risk review model is used to process the documents to be judged, and the risk review results are obtained, including:

[0042] The path length of the document to be judged in each isolated tree is calculated based on the feature vector corresponding to the document to be judged. The path length is used to represent the number of nodes that the document to be judged goes through from the root node of the isolated tree, selecting child nodes according to the feature vector corresponding to the document to be judged, until reaching the corresponding leaf node.

[0043] The risk review result is determined based on the average length of each path. The smaller the average length of each path, the greater the probability that the document to be judged is a false alarm.

[0044] According to another aspect of this disclosure, a risk review apparatus is provided. The apparatus includes:

[0045] The first determining module is used to perform feature extraction based on at least one document sample to obtain multiple types of feature vectors. The multiple types of feature vectors include at least two of text feature vectors, numerical feature vectors, and graph feature vectors. The at least one document sample does not include known false alarm samples or includes some known false alarm samples.

[0046] The second determination module is used to determine the risk review model based on the feature vectors of multiple types corresponding to each document sample;

[0047] The third determination module is used to process the documents to be judged based on the risk review model to obtain the risk review results. The risk review results are used to indicate whether there are false alarms in the documents to be judged.

[0048] In one possible implementation, the first determining module is used for:

[0049] Preprocess at least one text field in a document sample to obtain a preprocessed document sample.

[0050] Text enhancement processing is performed on the target document sample in the preprocessed document sample to obtain positive samples, and some or all of the samples are selected from other document samples as negative samples.

[0051] The target document sample, positive sample, and negative sample are encoded using a text encoder to obtain the original semantic vectors corresponding to the target document sample, positive sample, and negative sample, respectively.

[0052] The parameters of the text encoder are optimized and trained based on the original semantic vectors corresponding to the target document samples, positive samples, and negative samples, and the text feature vectors corresponding to each document sample are obtained based on the trained text encoder.

[0053] In one possible implementation, the parameters of the text encoder are optimized and trained based on the original semantic vectors corresponding to the target document samples, positive samples, and negative samples, respectively. The trained text encoder then yields the text feature vectors corresponding to each document sample, including:

[0054] The first loss function value is calculated based on the original semantic vectors corresponding to the target document sample, positive sample and negative sample respectively. The parameters of the text encoder are optimized and trained based on the first loss function value. The first semantic vector corresponding to each document sample is obtained using the trained text encoder.

[0055] The first semantic vector is processed by a multilayer perceptron encoder to obtain a first latent semantic vector. The dimension of the first latent semantic vector is smaller than that of the first semantic vector.

[0056] The first latent semantic vector is processed by a multilayer perceptron decoder to obtain a second semantic vector, which has the same dimension as the first semantic vector.

[0057] The second loss function value is calculated based on the first and second semantic vectors. The parameters of the multilayer perceptron encoder are optimized and trained based on the second loss function value. The second latent semantic vector corresponding to each document sample is obtained by using the trained multilayer perceptron encoder as the text feature vector.

[0058] In one possible implementation, the first determining module is used for:

[0059] Numerical fields in at least one document sample are preprocessed according to different types to obtain multiple types of numerical features;

[0060] For any type of numerical feature, it is grouped according to a preset grouping method to obtain multiple sets of numerical features corresponding to that type.

[0061] For any type of numerical feature, based on the proportion of false alarm samples and non-false alarm samples in the corresponding document samples of each group of numerical features, at least one type of numerical feature that meets the preset conditions is selected from multiple types as a numerical feature vector.

[0062] In one possible implementation, the first determining module is used for:

[0063] Entities and relationships are extracted from at least one document sample to determine the attribute values ​​of each entity and the relationships between entities, so as to obtain a knowledge graph. Nodes in the knowledge graph are used to represent the corresponding entities, and edges in the knowledge graph are used to represent the relationships between entities. Each node is associated with a category label, which is used to indicate the category of the entity corresponding to the node.

[0064] For any target node in the knowledge graph, a graph neural network is used to process the feature vector of the target node and the feature vectors of the target node's neighboring nodes to obtain the first multidimensional feature vector.

[0065] The predicted category of the entity corresponding to the target node is determined based on the first multidimensional feature vector, and the third loss function value is calculated based on the predicted category and the category label of the target node.

[0066] The parameters of the graph neural network are optimized and trained based on the value of the third loss function, and the second multidimensional feature vector corresponding to the knowledge graph is obtained by using the trained graph neural network.

[0067] For any document sample, the feature vectors of the nodes corresponding to any document sample in the second multidimensional feature vector are combined to form the graph feature vector corresponding to that document sample.

[0068] In one possible implementation, the second determining module is used for:

[0069] In the case where at least one document sample includes a known false alarm sample, at least one set of sample pairs and the label value corresponding to each sample pair are determined based on at least one document sample. The types of sample pairs include: a first sample pair consisting of two known false alarm samples, a second sample pair consisting of a known false alarm sample and an unknown sample, and a third sample pair consisting of two unknown samples. The label value of the first sample pair is greater than the label value of the second sample pair, and the label value of the second sample pair is greater than the label value of the third sample pair.

[0070] For any pair of samples, input the feature vectors corresponding to each document sample in the pair into the end-to-end learner to obtain the output vector;

[0071] The output vector is input into the ordinal regressor to obtain the predicted value;

[0072] The fourth loss function value is calculated based on the label value and the predicted value. The parameters of the end-to-end learner and the ordinal regressor are optimized and trained based on the fourth loss function value to obtain the risk review model.

[0073] In one possible implementation, the third determining module is used for:

[0074] From at least one document sample, some known false alarm samples are extracted and combined with the document to be judged to form multiple first sample pairs to be judged; and from at least one document sample, some unknown samples are extracted and combined with the document to be judged to form multiple second sample pairs to be judged.

[0075] Based on the risk review model, multiple first and second sample pairs to be discriminated are processed to obtain the corresponding first and second discriminant scores.

[0076] The risk review results are determined based on the first and second discriminant scores.

[0077] In one possible implementation, the second determining module is used for:

[0078] In the case that at least one document sample does not include known false alarm samples, an isolation tree is constructed based on the feature vectors of each type corresponding to each document sample, wherein each leaf node in the isolation tree corresponds to the feature vector of a document sample.

[0079] Use at least one isolated tree as the risk review model;

[0080] The third determining module is used for:

[0081] The path length of the document to be judged in each isolated tree is calculated based on the feature vector corresponding to the document to be judged. The path length is used to represent the number of nodes that the document to be judged goes through from the root node of the isolated tree, selecting child nodes according to the feature vector corresponding to the document to be judged, until reaching the corresponding leaf node.

[0082] The risk review result is determined based on the average length of each path. The smaller the average length of each path, the greater the probability that the document to be judged is a false alarm.

[0083] According to another aspect of this disclosure, a risk review apparatus is provided, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-described method.

[0084] According to another aspect of this disclosure, a non-volatile computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the above-described method.

[0085] According to another aspect of this disclosure, a computer program product is provided, including a computer program or a non-volatile computer-readable storage medium carrying the computer program, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0086] According to embodiments of this disclosure, feature vectors of at least two types, including textual feature vectors, numerical feature vectors, and graph feature vectors, are extracted based on at least one document sample. The at least one document sample does not include known false alarm samples or includes some known false alarm samples. Based on the multiple types of feature vectors corresponding to each document sample, a risk review model is determined. The document to be judged is processed based on the risk review model to obtain the risk review result. This effectively mines and utilizes valuable information from historical false alarm data. In scenarios lacking labeled samples, it fully utilizes the large amount of structured and unstructured information contained in the document data to more comprehensively and deeply understand the contextual semantics, statistical patterns, and entity relationships contained in the express delivery document information, thereby improving the ability to distinguish between potential risk features and false alarm signals. Furthermore, the solution of this disclosure, by extracting multiple types of feature vectors, can reduce dependence on a single data source or feature type, making the model more adaptable and capable of judgment when facing new and unseen document patterns, reducing the problem of decreased recognition accuracy due to changes in data distribution, and thus effectively reducing the false alarm rate.

[0087] Other features and aspects of this disclosure will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0088] The accompanying drawings, which are included in and form part of this specification, illustrate exemplary embodiments, features, and aspects of this disclosure together with the specification and serve to explain the principles of this disclosure.

[0089] Figure 1 shows a flowchart of a risk review method according to an embodiment of the present disclosure.

[0090] Figure 2 shows a flowchart of a risk review method according to an embodiment of the present disclosure.

[0091] Figure 3 illustrates a flowchart of determining text feature vectors according to an embodiment of the present disclosure.

[0092] Figure 4 shows a schematic diagram of extracting spectral feature vectors according to an embodiment of the present disclosure.

[0093] Figure 5 shows a schematic diagram of the structure of an ordinal regression global model according to an embodiment of the present disclosure.

[0094] Figure 6 shows a schematic diagram of an isolated tree according to an embodiment of the present disclosure.

[0095] Figure 7 shows a structural diagram of a risk review apparatus according to an embodiment of the present disclosure.

[0096] Figure 8 is a block diagram illustrating an apparatus 1900 for risk review according to an exemplary embodiment. Detailed Implementation

[0097] Various exemplary embodiments, features, and aspects of this disclosure will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0098] As used herein, the terms “comprising,” “including,” “having,” or variations thereof are open-ended and include one or more of the stated features, integrals, elements, steps, components, or functions, but do not exclude the presence or addition of one or more other features, integrals, elements, steps, components, functions, or groups thereof.

[0099] When an element is referred to as “connected,” “coupled,” “responding,” or a variation thereof relative to another element, it may be directly connected, coupled, or responding to another element, or there may be an intermediate element present.

[0100] Although the terms first, second, third, etc., may be used herein to describe various elements / operations, these elements / operations should not be limited by these terms. These terms are only used to distinguish one element / operation from another. Therefore, without departing from the teachings of the inventive concept, a first element / operation in some embodiments may be referred to as a second element / operation in other embodiments.

[0101] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.

[0102] Furthermore, to better illustrate this disclosure, numerous specific details are set forth in the following detailed description. Those skilled in the art will understand that this disclosure can be practiced without certain specific details. In some instances, methods, means, components, and circuits well known to those skilled in the art have not been described in detail in order to highlight the main points of this disclosure.

[0103] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, data stored, data displayed, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant regions.

[0104] In existing customs express clearance operations, the identification of prohibited and restricted items mainly relies on target detection algorithms based on intelligent image analysis. However, due to factors such as non-standard declaration data, high complexity of commodity forms, and insufficient algorithm generalization ability, these algorithms are prone to misreporting non-prohibited and restricted items as prohibited or restricted items. Such misreporting issues lead to an increase in the workload of manual verification by customs and a extension of clearance time. Therefore, how to effectively reduce the misreporting rate has become an urgent problem to be solved in the current customs supervision field.

[0105] Currently, research on risk verification of prohibited and restricted items mainly focuses on two aspects: hardware optimization, such as increasing the radiation intensity and resolution of equipment, adding multi-angle scanning, and using dynamic rotation imaging technology; and algorithmic improvement, such as increasing the diversity of training data, capturing more dimensional image features, and adopting more advanced algorithm network structures. However, the above optimization methods have certain limitations: firstly, they rely too heavily on image data and fail to fully utilize the large amount of structured and unstructured information contained in the document data, which may be related to false alarm behavior; secondly, in practical application scenarios, the number of false alarm samples is often limited and difficult to collect, which can lead to overfitting or underfitting problems in the model. Therefore, in scenarios lacking labeled samples, how to fully mine and utilize document information to construct an auxiliary false alarm risk analysis mechanism has become an important research direction for risk verification methods of prohibited and restricted items.

[0106] In view of this, this disclosure provides a risk review method, apparatus, and storage medium. The method of this disclosure extracts at least two types of feature vectors, including text feature vectors, numerical feature vectors, and graph feature vectors, based on at least one document sample. The at least one document sample does not include known false alarm samples or includes some known false alarm samples. Based on the multiple types of feature vectors corresponding to each document sample, a risk review model is determined. The document to be reviewed is processed based on the risk review model to obtain the risk review result. This method can effectively mine and utilize valuable information from historical false alarm data. In scenarios lacking labeled samples, it fully utilizes the large amount of structured and unstructured information contained in the document data to more comprehensively and deeply understand the contextual semantics, statistical patterns, and entity relationships contained in the express delivery document information, thereby improving the ability to distinguish between potential risk features and false alarm signals. Furthermore, by extracting multiple types of feature vectors, the solution of this disclosure can reduce dependence on a single data source or feature type, making the model more adaptable and capable of judgment when facing new and unseen document patterns, reducing the problem of decreased recognition accuracy due to changes in data distribution, and thus effectively reducing the false alarm rate.

[0107] This disclosure aims to reduce the false alarm rate of intelligent image review systems by utilizing historical express mail document information and false alarm data, analyzing document content, mining features, and building models. Embodiments of this disclosure can be applied to customs express mail clearance processes. The risk review model constructed in these embodiments can be deployed as a subsequent auxiliary module of the intelligent image review system within the customs express mail clearance process. For example, when the intelligent image review system performs initial screening of package images and generates a suspicious alert, the risk review model of this disclosure can further review the corresponding documents, thereby automatically filtering out a large number of non-risk alerts triggered by algorithm errors. This reduces the workload of manual review and improves the intelligence and efficiency of the customs clearance process.

[0108] The risk review method of this disclosure can be deployed on a terminal device or a server. The terminal device can be any one or more of the following: mobile phone, foldable electronic device, tablet computer, desktop computer, laptop computer, handheld computer, notebook computer, ultra-mobile personal computer (UMPC), netbook, cellular phone, personal digital assistant (PDA), and in-vehicle device, having wired or wireless communication capabilities. This disclosure does not impose any special limitations on the specific type of terminal device.

[0109] Servers can be located locally or in the cloud, and can be physical devices or virtual devices such as virtual machines and containers. They have wireless communication capabilities, which can be configured in the server's chip (system) or other components. Wireless communication capabilities can be implemented through mobile communication technologies such as 2G / 3G / 4G / 5G, as well as Wi-Fi, Bluetooth, frequency modulation (FM), data radio, and satellite communication; they can also communicate via wired connections to interact with other devices.

[0110] Figure 1 shows a flowchart of a risk review method according to an embodiment of the present disclosure. As shown in Figure 1, the method may include:

[0111] Step S101: Perform feature extraction based on at least one document sample to obtain feature vectors of multiple types.

[0112] The document sample can be a parcel document generated during the historical customs clearance process. It can include unstructured information such as the parcel's product name, sender and recipient information, and logistics information, as well as structured information such as the product amount, quantity, and size.

[0113] At least one document sample may not include known false alarm samples or may only include some known false alarm samples. Known false alarm samples are samples with labels, that is, samples that have been confirmed as false alarms in actual review (e.g., samples that were mistakenly judged by the intelligent drawing review system as carrying prohibited or restricted items, but were found to be not prohibited after verification).

[0114] Multiple types of feature vectors can include at least two of text feature vectors, numerical feature vectors, and graph feature vectors, which are used to represent the textual semantic features, statistical numerical features, and inter-entity relational structural features in the document, respectively, thereby realizing multi-dimensional modeling of document information.

[0115] Figure 2 shows a flowchart of the risk review method according to an embodiment of this disclosure. As shown in Figure 2, feature vectors of the above-mentioned multiple types can be extracted based on document samples ("express delivery documents" in the figure). The process of extracting feature vectors of multiple types is described below.

[0116] As shown in Figure 2, in the process of extracting text feature vectors, contrastive learning and autoencoders can be utilized. Contrastive learning is a learning method based on the similarity and differences between samples, which can effectively learn a general and robust representation of text. The high-dimensional text features obtained by contrastive learning are effectively reduced in dimensionality by the autoencoder, becoming dense and semantically rich low-dimensional text features (i.e., the text feature vectors in this disclosure). In the process of extracting numerical feature vectors, feature grouping and feature filtering can be performed. In the process of extracting graph feature vectors, a knowledge graph can be constructed, and graph representations can be extracted as graph feature vectors.

[0117] In one possible implementation, step S101 includes:

[0118] Preprocess the text fields in at least one document sample to obtain a preprocessed document sample; perform text augmentation on the target document sample in the preprocessed document sample to obtain a positive sample, and select some or all of the samples from other document samples as negative samples; use a text encoder to encode the target document sample, positive sample, and negative sample respectively to obtain the original semantic vectors corresponding to the target document sample, positive sample, and negative sample respectively; optimize and train the parameters of the text encoder based on the original semantic vectors corresponding to the target document sample, positive sample, and negative sample respectively, and obtain the text feature vector corresponding to each document sample based on the trained text encoder.

[0119] Text fields in the document sample may include product names, sender / recipient information, and logistics information. Preprocessing may include data cleaning of the text fields in the document sample. Data cleaning can be achieved using existing technologies, such as punctuation normalization, stop word removal, abnormal character filtering, and special character deletion to remove noise and structure the text data. After data cleaning, a BytePair Encoding (BPE) algorithm can be used to segment the text content, generating a text character sequence composed of sub-words or characters as the preprocessed document sample.

[0120] The target document samples can be randomly selected from the preprocessed document samples. Text augmentation of the target document samples can be achieved using existing techniques, such as random masking, random deletion, word order shuffling, and synonym replacement, to slightly perturb the text and generate positive samples that are semantically similar to the original text but different in form, thereby improving the model's robustness to text changes. The number of negative samples can be the same as or different from the number of positive samples. Preferably, the number of negative samples can be set to be greater than the number of positive samples to amplify the distinction between positive and negative samples in the semantic space, thereby enhancing the learning effect and improving the ability to discriminate feature vectors. For example, with 200 preprocessed document samples, 50 samples can be selected as target document samples, 50 positive samples can be generated based on these 50 samples, and 80 samples can be selected as negative samples from the remaining 150 samples.

[0121] A pre-trained language model based on the Transformer architecture (such as BERT) can be used as a text encoder to encode the target document sample, positive sample, and negative sample separately, so as to obtain the original semantic vectors corresponding to the target document sample, positive sample, and negative sample respectively. The encoding process can be represented as follows:

[0122] ; ; ;

[0123] in, This represents the target document sample. Indicates a positive sample. Indicates a negative sample. N is the number of samples. It can represent a text encoder. It can represent the original semantic vector corresponding to the target document sample. It can represent the original semantic vector corresponding to a positive sample. It can represent the original semantic vector corresponding to the negative sample.

[0124] In one possible implementation, during the process of optimizing and training the parameters of the text encoder based on the original semantic vectors corresponding to the target document samples, positive samples, and negative samples, and obtaining the text feature vectors corresponding to each document sample based on the trained text encoder, the following can be done:

[0125] The first loss function value is calculated based on the original semantic vectors corresponding to the target document sample, positive sample, and negative sample, respectively. The parameters of the text encoder are then optimized and trained based on the first loss function value. The trained text encoder is used to obtain the first semantic vector corresponding to each document sample. The first semantic vector is processed by a multilayer perceptron encoder to obtain a first latent semantic vector. The first latent semantic vector is then processed by a multilayer perceptron decoder to obtain a second semantic vector. The second loss function value is calculated based on the first and second semantic vectors. The parameters of the multilayer perceptron encoder are then optimized and trained based on the second loss function value. The trained multilayer perceptron encoder is used to obtain the second latent semantic vector corresponding to each document sample as the text feature vector.

[0126] To obtain a better semantic vector representation, contrastive learning can be used to optimize and train the text encoder, resulting in a vector representation that effectively characterizes the semantics of the document text. The first loss function value can be the loss function value from contrastive learning. This first loss function can be an information-noise contrastive estimation function, and its calculation method can be expressed as follows:

[0127] ;

[0128] in, This represents the loss function value of the i-th document sample. This represents a similarity function, which can be set to a pre-defined similarity level. This represents a pre-set temperature coefficient used to scale the cosine similarity. In the loss function described above, by increasing... To bring positive samples closer together, by reducing... This achieves the goal of pushing away negative samples. The optimization objective of the training is to minimize the value of the first loss function. During the optimization training process, optimization methods such as gradient descent can be used to optimize the text encoder. The parameters are updated until the value of the first loss function is less than a preset threshold, thus obtaining the trained text encoder.

[0129] Based on the trained text encoder, all preprocessed document samples can be encoded to generate the first semantic vector corresponding to each document sample.

[0130] Furthermore, the first semantic vectors obtained from the trained text encoder can be input into a multilayer perceptron (MLP) encoder for processing to obtain a first latent semantic vector. Subsequently, the first latent semantic vectors can be input into the MLP decoder to obtain a second semantic vector. The dimension of the first latent semantic vector is less than the dimension of the first semantic vector, while the dimension of the second semantic vector is the same as that of the first semantic vector.

[0131] The second loss function value can be calculated based on the first and second semantic vectors. For example, the reconstruction error can be used as the second loss function to optimize the parameters of the multilayer perceptron encoder. The calculation method of the second loss function value can be expressed as follows:

[0132] ;

[0133] in, This can represent the value of the second loss function. It can represent the first semantic vector. The second semantic vector can be represented by N, where N is the number of samples. The optimization objective of the training is to minimize the value of the second loss function. During the training process, the parameters of the multilayer perceptron encoder can be updated using optimization methods such as gradient descent until the value of the second loss function is less than a preset threshold, thus obtaining the trained multilayer perceptron encoder.

[0134] After optimization, each first semantic vector can be re-input into the trained multilayer perceptron encoder to obtain the second latent semantic vector corresponding to each document sample, which can be used as the text feature vector.

[0135] Through the above methods, on the one hand, contrastive learning is used to learn representations for the text encoder, enabling it to capture semantic differences in document text; on the other hand, the autoencoder structure is used to further extract compact and stable latent semantic features, so that the text features have better generalization ability and robustness in subsequent models.

[0136] Figure 3 illustrates a flowchart of determining text feature vectors according to an embodiment of the present disclosure. As shown in Figure 3, textual fields (such as product names) in the document sample can first be preprocessed. For example, for the textual field "10 shirts", the quantity and unit "10" can be removed, leaving "shirts" as the target document sample (as shown in the "text sample" after text preprocessing in the figure). Subsequently, text enhancement techniques can be used to generate multiple semantically similar samples, such as "shirt" and "cotton shirt", as positive samples; and multiple samples can be selected from the product names of other document samples as negative samples.

[0137] The target document sample and positive and negative samples can be encoded using a BERT encoder (as shown in the "text encoder" diagram) to obtain the original semantic vector (with a vector dimension of, for example, 768). The original semantic vectors corresponding to the target document sample and the positive and negative samples are then used to form sample pairs, which are fed into a contrastive learner for training. This results in higher semantic vector similarity between the target document sample and the positive sample, and lower semantic vector similarity between the target document sample and the negative sample. The parameters of the text encoder can be optimized using the above method. After training, the first semantic vector (with a vector dimension of, for example, 768) can be obtained using the trained text encoder.

[0138] Furthermore, the above method can be used to train an autoencoder (where the encoder dimension is, for example, 768×4, and the decoder dimension is, for example, 4×768). The first semantic vector can be input into the trained autoencoder, and after processing by the encoder part, a second latent semantic vector (with a dimension of, for example, 4) can be obtained, which serves as the final text feature vector corresponding to the product name. For other text-type fields, the same method can be used to process them, thereby obtaining the second latent semantic vector corresponding to each document sample.

[0139] In one possible implementation, step S101 includes:

[0140] Numerical fields in at least one document sample are preprocessed according to different types to obtain multiple types of numerical features; for any type of data feature, they are grouped according to a preset grouping method to obtain multiple sets of numerical features corresponding to that type; for any type of numerical feature, based on the ratio of false alarm samples and non-false alarm samples in the document sample corresponding to each set of numerical features, at least one type of numerical feature that meets the preset conditions is selected from multiple types as a numerical feature vector.

[0141] The above procedure can be applied to extract numerical feature vectors only if at least one document sample includes labeled samples (i.e., samples with known false alarms or non-false alarms). If at least one document sample does not include labeled samples, the above numerical feature vectors do not need to be extracted.

[0142] Numerical fields in a sample document may include product price, quantity, and size. Preprocessing can involve data cleaning of numerical fields to remove missing and outlier values ​​from the original data, ensuring consistency and accuracy. For example, fields with a missing value ratio exceeding a set threshold can be directly deleted.

[0143] Different data types can reflect the business meaning and distribution characteristics of numeric fields. This can be determined based on the field's position in the document sample. Different preprocessing methods can be used for different types of numeric fields. For example, missing values ​​in numeric fields such as weight and size can be filled using methods like mean or median. Existing technologies such as box plots can be used to identify data that clearly exceeds reasonable business scope (e.g., excessively large volumes) and delete it. For date units, weight units, and currency units, different units can be standardized to a uniform format according to preset rules or mapping tables.

[0144] For example, when cleaning data for the product amount field, the following processing can be performed: samples with negative amounts or exceeding the business limit (e.g., 5,000 yuan) can be directly deleted; for data with missing amount fields, the average value of the same product name in other document samples can be used to fill the missing amount; for amount fields using different currency units (e.g., RMB, USD, EUR, JPY, etc.), they can be uniformly converted to the same currency unit based on the preset exchange rate.

[0145] Preset grouping methods include fixed-range grouping and equal-frequency grouping. Fixed-range grouping divides the value range of a numeric field into groups of equal width. For example, for fields like product price, fixed-range grouping can be used, grouping the price range into groups such as 0–1000, 1000–2000, etc. Equal-frequency grouping sorts the samples by value so that each group contains the same number of samples. For example, equal-frequency grouping can be used for fields like date. Through these grouping methods, multiple sets of numeric features corresponding to the respective type can be obtained.

[0146] Specifically, using labeled samples from at least one document sample, at least one type of numerical feature that meets preset conditions is selected through the following process, and then the at least one type of numerical feature selected above is extracted as a numerical feature vector for unlabeled samples from at least one document sample.

[0147] For each set of numerical features with labeled samples, the weight of evidence (WOE) can be calculated based on the proportion of false positives and non-false positives in the corresponding document samples within each set of numerical features. The calculation method can be expressed as follows:

[0148] ;

[0149] in, This represents the WOE value corresponding to the m-th group. Indicates the first The number of non-false alarm samples in the group This represents the number of all non-false alarm samples; Indicates the first The number of false positives in the group This indicates the total number of false positives. This represents the proportion of non-false alarm samples in group m out of the total number of non-false alarm samples. This represents the proportion of false positive samples in group m out of the total number of false positive samples.

[0150] Furthermore, based on the WOE value, the Information Value (IV) index can be calculated, and its calculation method can be expressed as follows:

[0151] ;

[0152] in, The IV value indicates the number of groups after the variable is grouped. The larger the IV value, the stronger the predictive ability of the feature, that is, the extent to which the corresponding numerical feature (such as amount or date) can distinguish between false positive samples (Bad) and non-false positive samples (Good).

[0153] Preset conditions can include that the IV value of the numerical feature of this type is greater than the IV threshold. The IV threshold can be preset according to empirical rules and actual business needs. For example, it can be set to 0.3. An IV value less than 0.02 can indicate that the corresponding type of numerical feature has extremely weak predictive ability and is usually rejected. An IV value greater than 0.3 can indicate that the corresponding type of numerical feature has strong predictive ability. When the IV value is greater than or equal to 0.5, it is also necessary to check whether overfitting is possible.

[0154] For example, for numerical features such as the price of goods, the product can be divided into five equal intervals based on its price. The number of false positives and true negatives within each interval can be counted. Then, according to the calculation method described above, the WOE and IV values ​​for each group can be calculated, as shown in Table 1.

[0155] Table 1

[0156]

[0157] Based on the IV values ​​of each group in Table 1, the overall IV value for the numerical feature of commodity amount is 0.31, indicating that this feature has a strong predictive ability and should be retained.

[0158] Therefore, by filtering numerical features, we can not only reduce modeling complexity and avoid the risk of the curse of dimensionality, but also reduce subjective interference from human experience and enhance the interpretability and comparability of the results.

[0159] In one possible implementation, step S101 includes:

[0160] Entities and relationships are extracted from at least one document sample to determine the attribute values ​​of each entity and the relationships between entities, thus obtaining a knowledge graph. For any target node in the knowledge graph, a graph neural network is used to process the feature vector of the target node and the feature vectors of its neighboring nodes to obtain a first multidimensional feature vector. Based on the first multidimensional feature vector, the predicted category of the entity corresponding to the target node is determined. Based on the predicted category and the category label of the target node, a third loss function value is calculated. Based on the third loss function value, the parameters of the graph neural network are optimized and trained, and the trained graph neural network is used to obtain the second multidimensional feature vector corresponding to the knowledge graph. For any document sample, the feature vectors of the nodes corresponding to any document sample in the second multidimensional feature vector are combined to obtain the graph feature vector corresponding to that document sample.

[0161] First, the categories and attributes of entities in the knowledge graph, as well as the categories and attributes of relationships between entities, can be defined to facilitate subsequent aggregation processing using a consistent graph neural network model.

[0162] Figure 4 illustrates a schematic diagram of extracting graph feature vectors according to an embodiment of the present disclosure. As shown on the left side of Figure 4, nodes in the knowledge graph can be used to represent corresponding entities, and edges in the knowledge graph can be used to represent relationships between entities. The categories of entities can include, for example, recipients, senders, and goods. Each type of entity can further include attributes such as name, address, and product type. Relationships between entities can include, for example, sending / receiving, receiving goods, and shipping goods. The attributes of sending / receiving can include the number of sending / receiving transactions and the number of false alarms. The attributes of receiving goods can include the number of receiving transactions and the number of false alarms. The attributes of shipping goods can include the number of shipping transactions and the number of false alarms.

[0163] Each node in the knowledge graph can be associated with a category label. The category label is used to indicate the category of the entity corresponding to the node, such as indicating that the entity is a recipient, a shipper, or a product category.

[0164] You can select some fields (such as recipient, sender, product, etc.) from the document sample as the data source for building the knowledge graph, and clean the data source, including noise reduction, deduplication, and format unification. The cleaning methods for text fields and numerical fields can be performed in accordance with the aforementioned steps.

[0165] Based on the cleaned data, relevant entity and relationship instances can be automatically extracted according to predefined rules or structured modules, and attribute values ​​of various entities and relationships can be statistically analyzed. Furthermore, text similarity-based calculation methods can be used to merge multiple entities that may have the same name or alias, eliminating redundant nodes and ensuring the uniqueness of entity referencing.

[0166] For example, for a node with the product name "xx banana wafer biscuits", the product name can be denoised and standardized to "wafer biscuits". For different nodes with recipient names "xxx Co., Ltd." and "xxx Limited Liability Company", their text similarity is calculated to be 0.9 (greater than the preset similarity threshold). Entity alignment can be performed on these two nodes and they can be merged. Thus, the knowledge graph shown in the upper right corner of Figure 4 can be obtained. Based on this, the graph neural network can be trained using the Graph Sampling and Aggregate (GraphSAGE) algorithm to obtain the second multi-dimensional feature vector (the "graph representation vector" in Figure 4).

[0167] During training, you can first initialize the feature vectors of all nodes in the knowledge graph, set the number of aggregation layers of the graph neural network (e.g., 2), set the aggregation function AGG (Aggregation) to mean aggregation, and initialize the weight matrix.

[0168] For any target node in a knowledge graph, its neighbor nodes are the nodes directly connected to that target node in the knowledge graph. A neighbor set can be obtained by uniformly sampling from its neighbor nodes. The feature vectors of each neighbor node in the neighbor set can be aggregated, and then the feature vector of the target node can be combined with the aggregated feature vector through an activation function to obtain the embedding representation of the target node as the first multi-dimensional feature vector. The calculation method of the k-th layer embedding representation can be expressed as:

[0169] ;

[0170] in, Let v be the embedding representation of the target node at the k-th layer. This represents the activation function. It is the mean aggregation function. Let be the weight matrix of the k-th layer. The set of neighbors of the target node The identifier of the middle neighbor node. (After...) After the layer network, the final embedding vector of the target node can be obtained. Then, it is mapped into a first multidimensional feature vector through a multilayer perceptron. (Its dimension can be the same as the number of label categories of the node).

[0171] A normalization function (such as the softmax function) can be used to normalize the first multidimensional feature vector to obtain the predicted category of the entity corresponding to the target node. The category label classification of the node can be selected as the training objective. The third loss function can be the cross-entropy function, and the calculation method for the third loss function value can be expressed as:

[0172] ;

[0173] Where Loss represents the value of the third loss function. This represents the proportion of samples corresponding to the p-th category label to the total number of samples. This represents the actual category label of the entity corresponding to the target node. This indicates the predicted category of the entity corresponding to the target node.

[0174] The optimization objective of the training is to minimize the value of the third loss function. During the optimization training process, the parameters of the graph neural network can be updated using optimization methods such as gradient descent until the value of the third loss function is less than a preset threshold, resulting in the trained graph neural network. After optimization, the first multidimensional feature vectors corresponding to each node can be recalculated using the trained graph neural network in the manner described above. The first multidimensional feature vectors corresponding to each node are then combined to obtain the second multidimensional feature vector corresponding to the knowledge graph.

[0175] For any document sample, we can find the vectors corresponding to all entity nodes (such as sender nodes, recipient nodes, and product nodes) associated with the document sample in the second multidimensional feature vector, and obtain the graph feature vector corresponding to the document sample by concatenating them.

[0176] Step S102: Determine the risk review model based on the feature vectors of multiple types corresponding to each document sample.

[0177] The feature vectors of multiple types corresponding to each document sample can be concatenated to obtain a fused feature vector for each document sample.

[0178] Referring to Figure 2, in this embodiment of the disclosure, different risk review models can be constructed for weakly supervised scenarios (i.e., the document samples include some known false alarm samples) and unsupervised scenarios (i.e., the document samples do not include known false alarm samples).

[0179] Step S103: Process the documents to be judged based on the risk review model to obtain the risk review results.

[0180] The risk review results are used to indicate whether there is a false alarm in the document to be judged, that is, to indicate whether the document to be judged has been mistakenly judged as carrying prohibited or restricted items.

[0181] According to embodiments of this disclosure, feature vectors of at least two types, including textual feature vectors, numerical feature vectors, and graph feature vectors, are extracted based on at least one document sample. The at least one document sample does not include known false alarm samples or includes some known false alarm samples. Based on the multiple types of feature vectors corresponding to each document sample, a risk review model is determined. The document to be judged is processed based on the risk review model to obtain the risk review result. This effectively mines and utilizes valuable information from historical false alarm data. In scenarios lacking labeled samples, it fully utilizes the large amount of structured and unstructured information contained in the document data to more comprehensively and deeply understand the contextual semantics, statistical patterns, and entity relationships contained in the express delivery document information, thereby improving the ability to distinguish between potential risk features and false alarm signals. Furthermore, the solution of this disclosure, by extracting multiple types of feature vectors, can reduce dependence on a single data source or feature type, making the model more adaptable and capable of judgment when facing new and unseen document patterns, reducing the problem of decreased recognition accuracy due to changes in data distribution, and thus effectively reducing the false alarm rate.

[0182] The following sections will introduce the weakly supervised and unsupervised scenarios separately.

[0183] In one possible implementation, for weakly supervised scenarios, step S102 above may include:

[0184] If at least one document sample includes a known false alarm sample, at least one set of sample pairs and the corresponding label values ​​for each sample pair are determined based on at least one document sample. For any set of sample pairs, the feature vectors corresponding to each document sample in the sample pair are input into an end-to-end learner to obtain an output vector. The output vector is input into an ordinal regressor to obtain a predicted value. A fourth loss function value is calculated based on the label value and the predicted value. The parameters of the ordinal regressor are optimized and trained based on the fourth loss function value to obtain a risk review model.

[0185] The types of sample pairs can include: a first sample pair consisting of two known false positive samples, a second sample pair consisting of a known false positive sample and an unknown sample, and a third sample pair consisting of two unknown samples. Different label values ​​can be preset for different types of sample pairs, with the label value of the first sample pair being greater than that of the second sample pair, and the label value of the second sample pair being greater than that of the third sample pair.

[0186] Figure 5 shows a schematic diagram of the structure of an ordinal regression global model according to an embodiment of the present disclosure. As shown in Figure 5, the ordinal regression global model may include an end-to-end learner and an ordinal regressor.

[0187] end-to-end learner It can consist of two branches, with both branches having the same network structure (e.g., a fully connected network that shares parameters). The input to each branch is the fused feature vector of one sample from the sample pair. Its output vector z can be expressed as: .

[0188] Ordinal Regressor It can be a single-layer fully connected network, with the following parameters: Used to connect end-to-end learners The output vector Mapped to predicted values , can be represented as:

[0189] ;

[0190] in, It can be used to represent the fused feature vector corresponding to the input sample pair. and These represent end-to-end learners. For the sample and samples The output vector.

[0191] The fourth loss function can be the mean absolute error, and its value can be expressed as follows:

[0192] ;

[0193] in, express The corresponding fourth loss function value, It can represent The corresponding tag value.

[0194] The optimization objective of the training is to minimize the value of the fourth loss function. During the training process, optimization methods such as gradient descent can be used to adjust the parameters. (Including shared parameters of end-to-end learners) and the parameters of the ordinal regressor The process is repeated until the value of the fourth loss function is less than the preset threshold, resulting in the trained end-to-end learner and the trained ordinal regressor as the risk review model.

[0195] In the process of false alarm detection, step S103 may include:

[0196] Multiple first-to-be-discriminated sample pairs are formed by extracting some known false alarm samples from at least one document sample and combining them with the document to be discriminated. Multiple second-to-be-discriminated sample pairs are formed by extracting some unknown samples from at least one document sample and combining them with the document to be discriminated. The multiple first-to-be-discriminated sample pairs and second-to-be-discriminated sample pairs are processed based on the risk review model to obtain the corresponding first-to-be-discriminated score and second-to-be-discriminated score. The risk review result is determined based on the first-to-be-discriminated score and the second-to-be-discriminated score.

[0197] The document to be judged can be a document that has been initially screened by the intelligent drawing review system as potentially carrying prohibited or restricted items. Based on the process in step S101 above, the feature vectors of each type corresponding to the document to be judged can be determined, and then concatenated to obtain the fused feature vector corresponding to the document to be judged.

[0198] Some known false alarm samples can be M samples (e.g., 10) drawn from known false alarm samples, and some unknown samples can be M samples drawn from unknown samples. The fused feature vectors corresponding to each first pair of samples to be judged (including the fused feature vectors corresponding to the known false alarm samples and the fused feature vectors corresponding to the documents to be judged) and each second pair of samples to be judged (including the fused feature vectors corresponding to the unknown samples and the fused feature vectors corresponding to the documents to be judged) are respectively input into the trained end-to-end learner and the trained ordinal regressor. The process is described above to calculate the risk review score, which can be expressed as follows:

[0199] ;

[0200] in, It can represent the fused feature vector corresponding to the document to be judged. This indicates the risk review score corresponding to the document to be judged. This represents the parameters of the trained end-to-end learner and the trained ordinal regressor. This represents the fused feature vector corresponding to a known false positive sample. This can represent the first pair of samples to be discriminated in the i-th group. This can represent the first discrimination score corresponding to the first sample pair to be discriminated in the i-th group; This represents the fused feature vector corresponding to the unknown sample. This can represent the second pair of samples to be discriminated in the i-th group. It can represent the second discrimination score corresponding to the second sample pair to be discriminated in the i-th group.

[0201] The risk review score can be used to indicate the probability that the document to be judged is a false alarm. The higher the risk review score, the greater the probability that the document to be judged is a false alarm.

[0202] In one possible implementation, for unsupervised scenarios, step S102 includes:

[0203] If at least one document sample does not include known false alarm samples, an isolation tree is constructed based on the feature vectors of each type corresponding to each document sample; at least one isolation tree is used as the risk review model.

[0204] Specifically, a predetermined number of samples can be randomly selected from at least one document sample as subsamples. An isolated tree is built based on the feature vectors of each type corresponding to the subsamples. Then, new subsamples are randomly selected from the at least one document sample again, and another isolated tree is built based on the new subsamples. This process is repeated until the number of isolated trees reaches a predetermined threshold (e.g., 100 trees), thereby forming an isolated forest (i.e., including at least one isolated tree), which is used as a risk verification model. Figure 6 shows a schematic diagram of an isolated tree according to an embodiment of this disclosure. n isolated trees (isolated tree 1 - isolated tree n) as shown in Figure 6 can be built.

[0205] An isolation tree can be a binary search tree. During the construction of the isolation tree for each type of feature vector corresponding to a subsample, all subsamples can be placed at the root node of the isolation tree, which is the initial node used to begin recursive sample partitioning. For any node in the isolation tree, a feature vector of a certain type can be randomly selected, and a split value can be randomly determined within the range of values ​​for that feature vector. Then, the samples in that node can be divided into left and right child nodes based on the split value. The left child node contains samples whose feature vectors are less than the split value, and the right child node contains samples whose feature vectors are greater than or equal to the split value. The above steps are repeated for the left and right child nodes until a stopping condition is met (e.g., the node contains only one sample and cannot be further partitioned). At this point, the node is marked as a leaf node. A leaf node represents the terminating node when the isolation tree construction ends, and each leaf node in the isolation tree corresponds to one document sample.

[0206] In one possible implementation, step S103 includes:

[0207] The path length of the document to be judged in each isolated tree is calculated based on the feature vector corresponding to the document to be judged; the risk review result is determined based on the average value of each path length.

[0208] The path length represents the number of nodes traversed from the root node of the isolated tree, selecting child nodes based on the feature vector corresponding to the document, until reaching the corresponding leaf node. The smaller the average path length, the higher the probability that the document is a false alarm. A false alarm score can be calculated based on the average path length, which can be expressed as follows:

[0209] ;

[0210] in, This represents the false alarm score corresponding to the document to be judged, where x is the document to be judged. This represents the path length of x in an isolated tree. Let x be the average path length of x across all isolated trees. This represents the average path length for constructing a binary search tree given a sample size N. This value is a known fixed value.

[0211] in, The closer the value is to 1, the greater the probability that the document to be judged is a false alarm; the closer the value is to 0.5, the smaller the probability that the document to be judged is a false alarm.

[0212] Using the above methods, isolated forest models can be used to automatically identify documents with abnormal characteristics in unsupervised scenarios, thereby completing the false alarm risk review of the documents to be judged.

[0213] Figure 7 shows a structural diagram of a risk review apparatus according to an embodiment of the present disclosure. As shown in Figure 7, the apparatus may include:

[0214] The first determining module 701 is used to perform feature extraction based on at least one document sample to obtain multiple types of feature vectors. The multiple types of feature vectors include at least two of text feature vectors, numerical feature vectors, and graph feature vectors. The at least one document sample does not include known false alarm samples or includes some known false alarm samples.

[0215] The second determination module 702 is used to determine the risk review model based on the feature vectors of multiple types corresponding to each document sample;

[0216] The third determination module 703 is used to process the document to be judged based on the risk review model to obtain the risk review result. The risk review result is used to indicate whether there is a false alarm in the document to be judged.

[0217] In one possible implementation, the first determining module 701 is used for:

[0218] Preprocess at least one text field in a document sample to obtain a preprocessed document sample.

[0219] Text enhancement processing is performed on the target document sample in the preprocessed document sample to obtain positive samples, and some or all of the samples are selected from other document samples as negative samples.

[0220] The target document sample, positive sample, and negative sample are encoded using a text encoder to obtain the original semantic vectors corresponding to the target document sample, positive sample, and negative sample, respectively.

[0221] The parameters of the text encoder are optimized and trained based on the original semantic vectors corresponding to the target document samples, positive samples, and negative samples, and the text feature vectors corresponding to each document sample are obtained based on the trained text encoder.

[0222] In one possible implementation, the parameters of the text encoder are optimized and trained based on the original semantic vectors corresponding to the target document samples, positive samples, and negative samples, respectively. The trained text encoder then yields the text feature vectors corresponding to each document sample, including:

[0223] The first loss function value is calculated based on the original semantic vectors corresponding to the target document sample, positive sample and negative sample respectively. The parameters of the text encoder are optimized and trained based on the first loss function value. The first semantic vector corresponding to each document sample is obtained using the trained text encoder.

[0224] The first semantic vector is processed by a multilayer perceptron encoder to obtain a first latent semantic vector. The dimension of the first latent semantic vector is smaller than that of the first semantic vector.

[0225] The first latent semantic vector is processed by a multilayer perceptron decoder to obtain a second semantic vector, which has the same dimension as the first semantic vector.

[0226] The second loss function value is calculated based on the first and second semantic vectors. The parameters of the multilayer perceptron encoder are optimized and trained based on the second loss function value. The second latent semantic vector corresponding to each document sample is obtained by using the trained multilayer perceptron encoder as the text feature vector.

[0227] In one possible implementation, the first determining module 701 is used for:

[0228] Numerical fields in at least one document sample are preprocessed according to different types to obtain multiple types of numerical features;

[0229] For any type of numerical feature, it is grouped according to a preset grouping method to obtain multiple sets of numerical features corresponding to that type.

[0230] For any type of numerical feature, based on the proportion of false alarm samples and non-false alarm samples in the corresponding document samples of each group of numerical features, at least one type of numerical feature that meets the preset conditions is selected from multiple types as a numerical feature vector.

[0231] In one possible implementation, the first determining module 701 is used for:

[0232] Entities and relationships are extracted from at least one document sample to determine the attribute values ​​of each entity and the relationships between entities, so as to obtain a knowledge graph. Nodes in the knowledge graph are used to represent the corresponding entities, and edges in the knowledge graph are used to represent the relationships between entities. Each node is associated with a category label, which is used to indicate the category of the entity corresponding to the node.

[0233] For any target node in the knowledge graph, a graph neural network is used to process the feature vector of the target node and the feature vectors of the target node's neighboring nodes to obtain the first multidimensional feature vector.

[0234] The predicted category of the entity corresponding to the target node is determined based on the first multidimensional feature vector, and the third loss function value is calculated based on the predicted category and the category label of the target node.

[0235] The parameters of the graph neural network are optimized and trained based on the value of the third loss function, and the second multidimensional feature vector corresponding to the knowledge graph is obtained by using the trained graph neural network.

[0236] For any document sample, the feature vectors of the nodes corresponding to any document sample in the second multidimensional feature vector are combined to form the graph feature vector corresponding to that document sample.

[0237] In one possible implementation, the second determining module 702 is used for:

[0238] In the case where at least one document sample includes a known false alarm sample, at least one set of sample pairs and the label value corresponding to each sample pair are determined based on at least one document sample. The types of sample pairs include: a first sample pair consisting of two known false alarm samples, a second sample pair consisting of a known false alarm sample and an unknown sample, and a third sample pair consisting of two unknown samples. The label value of the first sample pair is greater than the label value of the second sample pair, and the label value of the second sample pair is greater than the label value of the third sample pair.

[0239] For any pair of samples, input the feature vectors corresponding to each document sample in the pair into the end-to-end learner to obtain the output vector;

[0240] The output vector is input into the ordinal regressor to obtain the predicted value;

[0241] The fourth loss function value is calculated based on the label value and the predicted value. The parameters of the end-to-end learner and the ordinal regressor are optimized and trained based on the fourth loss function value to obtain the risk review model.

[0242] In one possible implementation, the third determining module 703 is used for:

[0243] From at least one document sample, some known false alarm samples are extracted and combined with the document to be judged to form multiple first sample pairs to be judged; and from at least one document sample, some unknown samples are extracted and combined with the document to be judged to form multiple second sample pairs to be judged.

[0244] Based on the risk review model, multiple first and second sample pairs to be discriminated are processed to obtain the corresponding first and second discriminant scores.

[0245] The risk review results are determined based on the first and second discriminant scores.

[0246] In one possible implementation, the second determining module 702 is used for:

[0247] In the case that at least one document sample does not include known false alarm samples, an isolation tree is constructed based on the feature vectors of each type corresponding to each document sample, wherein each leaf node in the isolation tree corresponds to the feature vector of a document sample.

[0248] Use at least one isolated tree as the risk review model;

[0249] The third determining module 703 is used for:

[0250] The path length of the document to be judged in each isolated tree is calculated based on the feature vector corresponding to the document to be judged. The path length is used to represent the number of nodes that the document to be judged goes through from the root node of the isolated tree, selecting child nodes according to the feature vector corresponding to the document to be judged, until reaching the corresponding leaf node.

[0251] The risk review result is determined based on the average length of each path. The smaller the average length of each path, the greater the probability that the document to be judged is a false alarm.

[0252] According to embodiments of this disclosure, feature vectors of at least two types, including textual feature vectors, numerical feature vectors, and graph feature vectors, are extracted based on at least one document sample. The at least one document sample does not include known false alarm samples or includes some known false alarm samples. Based on the multiple types of feature vectors corresponding to each document sample, a risk review model is determined. The document to be judged is processed based on the risk review model to obtain the risk review result. This effectively mines and utilizes valuable information from historical false alarm data. In scenarios lacking labeled samples, it fully utilizes the large amount of structured and unstructured information contained in the document data to more comprehensively and deeply understand the contextual semantics, statistical patterns, and entity relationships contained in the express delivery document information, thereby improving the ability to distinguish between potential risk features and false alarm signals. Furthermore, the solution of this disclosure, by extracting multiple types of feature vectors, can reduce dependence on a single data source or feature type, making the model more adaptable and capable of judgment when facing new and unseen document patterns, reducing the problem of decreased recognition accuracy due to changes in data distribution, and thus effectively reducing the false alarm rate.

[0253] In some embodiments, the functions or modules of the apparatus provided in this disclosure can be used to perform the methods described in the above method embodiments. The specific implementation can be referred to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0254] This disclosure also provides a risk review apparatus, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above method.

[0255] This disclosure also provides a non-volatile computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described method.

[0256] This disclosure also provides a computer program product, including a computer program or a non-volatile computer-readable storage medium carrying the computer program, wherein the computer program, when executed by a processor, implements the steps of the above method.

[0257] Figure 8 is a block diagram illustrating an apparatus 1900 for risk review according to an exemplary embodiment. For example, apparatus 1900 may be provided as a server or terminal device. Referring to Figure 8, apparatus 1900 includes a processing component 1922, which further includes one or more processors, and memory resources represented by memory 1932 for storing instructions executable by the processing component 1922, such as application programs. The application programs stored in memory 1932 may include one or more modules, each corresponding to a set of instructions. Furthermore, processing component 1922 is configured to execute instructions to perform the methods described above.

[0258] Device 1900 may also include a power supply component 1926 configured to perform power management of device 1900, a wired or wireless network interface 1950 configured to connect device 1900 to a network, and an input / output interface 1958 (I / O interface). Device 1900 can operate on an operating system, such as Windows Server, stored in memory 1932. TM macOS X TM Unix TM Linux TM FreeBSD TM Or similar.

[0259] In an exemplary embodiment, a non-volatile computer-readable storage medium is also provided, such as a memory 1932 including computer program instructions that can be executed by a processing component 1922 of the device 1900 to perform the above-described method.

[0260] Computer-readable storage media can be tangible devices capable of holding and storing programs / instructions used by instruction execution devices. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0261] The computer program (or computer-readable program instructions) described herein can be downloaded from a computer-readable storage medium to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage medium in the respective computing / processing device.

[0262] The computer program (or computer program instructions) used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing state information from the computer-readable program instructions to implement various aspects of this disclosure.

[0263] Various aspects of this disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0264] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0265] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0266] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0267] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A risk review method, characterized in that, The method includes: extracting features based on at least one document sample to obtain multiple types of feature vectors, wherein the multiple types of feature vectors include at least two of text feature vectors, numerical feature vectors, and graph feature vectors, and wherein the at least one document sample does not include known false alarm samples or includes some known false alarm samples; determining a risk review model based on the multiple types of feature vectors corresponding to each document sample; and processing the document to be judged based on the risk review model to obtain a risk review result, wherein the risk review result is used to indicate whether the document to be judged has false alarms.

2. The method according to claim 1, characterized in that, The step of extracting features based on at least one document sample to obtain multiple types of feature vectors includes: preprocessing text fields in at least one document sample to obtain preprocessed document samples; performing text enhancement processing on target document samples in the preprocessed document samples to obtain positive samples, and selecting some or all samples from other document samples as negative samples; encoding the target document sample, the positive samples, and the negative samples respectively using a text encoder to obtain the original semantic vectors corresponding to the target document sample, the positive samples, and the negative samples respectively; optimizing and training the parameters of the text encoder based on the original semantic vectors corresponding to the target document sample, the positive samples, and the negative samples respectively, and obtaining the text feature vectors corresponding to each document sample based on the trained text encoder.

3. The method according to claim 2, characterized in that, The step of optimizing and training the parameters of the text encoder based on the original semantic vectors corresponding to the target document sample, the positive sample, and the negative sample, and obtaining the text feature vector corresponding to each document sample based on the trained text encoder, includes: calculating a first loss function value based on the original semantic vectors corresponding to the target document sample, the positive sample, and the negative sample; optimizing and training the parameters of the text encoder based on the first loss function value; obtaining a first semantic vector corresponding to each document sample using the trained text encoder; processing the first semantic vector using a multilayer perceptron encoder to obtain a first latent semantic vector, the dimension of which is less than the dimension of the first semantic vector; processing the first latent semantic vector using a multilayer perceptron decoder to obtain a second semantic vector, the second semantic vector having the same dimension as the first semantic vector; calculating a second loss function value based on the first semantic vector and the second semantic vector; optimizing and training the parameters of the multilayer perceptron encoder based on the second loss function value; and using the second latent semantic vector corresponding to each document sample obtained by the trained multilayer perceptron encoder as the text feature vector.

4. The method according to claim 1, characterized in that, The step of extracting features based on at least one document sample to obtain multiple types of feature vectors includes: preprocessing the numerical fields in the at least one document sample according to different types to obtain multiple types of numerical features; grouping the numerical features of any type according to a preset grouping method to obtain multiple groups of numerical features corresponding to that type; and selecting at least one type of numerical features that meets preset conditions from the multiple types as the numerical feature vector based on the ratio of false alarm samples and non-false alarm samples in the document samples corresponding to each group of numerical features.

5. The method according to claim 1, characterized in that, The step of feature extraction based on at least one document sample to obtain multiple types of feature vectors includes: extracting entities and relationships from at least one document sample, determining the attribute values ​​of each entity and the relationships between entities to obtain a knowledge graph, wherein nodes in the knowledge graph represent corresponding entities, edges in the knowledge graph represent relationships between entities, each node is associated with a category label, and the category label indicates the category of the entity corresponding to the node; for any target node in the knowledge graph, processing the feature vector of the target node and the feature vectors of the target node's neighboring nodes using a graph neural network to obtain a first multidimensional feature vector; determining the predicted category of the entity corresponding to the target node based on the first multidimensional feature vector, calculating a third loss function value based on the predicted category and the category label of the target node; optimizing and training the parameters of the graph neural network based on the third loss function value, and obtaining a second multidimensional feature vector corresponding to the knowledge graph using the trained graph neural network; for any document sample, combining the feature vectors of the nodes corresponding to the document sample in the second multidimensional feature vector as the graph feature vector corresponding to that document sample.

6. The method according to claim 1, characterized in that, The step of determining the risk review model based on the feature vectors of the multiple types corresponding to each document sample includes: when the at least one document sample includes known false alarm samples, determining at least one set of sample pairs and the label value corresponding to each sample pair based on the at least one document sample, wherein the types of the sample pairs include: a first sample pair consisting of two known false alarm samples, a second sample pair consisting of a known false alarm sample and an unknown sample, and a third sample pair consisting of two unknown samples, wherein the label value of the first sample pair is greater than the label value of the second sample pair, and the label value of the second sample pair is greater than the label value of the third sample pair; for any sample pair, inputting the feature vectors corresponding to each document sample in the sample pair into an end-to-end learner to obtain an output vector; inputting the output vector into an ordinal regressor to obtain a predicted value; calculating a fourth loss function value based on the label value and the predicted value, and optimizing and training the parameters of the end-to-end learner and the ordinal regressor based on the fourth loss function value to obtain the risk review model.

7. The method according to claim 6, characterized in that, The step of processing the documents to be judged based on the risk review model to obtain the risk review result includes: extracting some known false alarm samples from the at least one document sample and forming multiple first sample pairs to be judged with the documents to be judged; and extracting some unknown samples from the at least one document sample and forming multiple second sample pairs to be judged with the documents to be judged; processing the multiple first sample pairs to be judged and the second sample pairs to be judged based on the risk review model to obtain corresponding first judgment scores and second judgment scores; and determining the risk review result based on the first judgment scores and the second judgment scores.

8. The method according to claim 1, characterized in that, The step of determining the risk review model based on the feature vectors of the multiple types corresponding to each document sample includes: when the at least one document sample does not include known false alarm samples, constructing isolated trees based on the feature vectors of each type corresponding to each document sample, wherein each leaf node in the isolated tree corresponds to a document sample; using at least one isolated tree as the risk review model; and processing the document to be judged based on the risk review model to obtain the risk review result includes: calculating the path length of the document to be judged in each isolated tree based on the feature vector corresponding to the document to be judged, wherein the path length is used to characterize the number of nodes traversed by the document to be judged from the root node of the isolated tree, selecting child nodes according to the feature vector corresponding to the document to be judged, until reaching the corresponding leaf node; and determining the risk review result based on the average value of each path length, wherein the smaller the average value of each path length, the greater the probability that the document to be judged is a false alarm.

9. A risk review device, characterized in that, The device includes: a first determining module, configured to extract features based on at least one document sample to obtain multiple types of feature vectors, wherein the multiple types of feature vectors include at least two of text feature vectors, numerical feature vectors, and graph feature vectors, and the at least one document sample does not include known false alarm samples or includes some known false alarm samples; a second determining module, configured to determine a risk review model based on the multiple types of feature vectors corresponding to each document sample; and a third determining module, configured to process the document to be judged based on the risk review model to obtain a risk review result, wherein the risk review result is used to indicate whether the document to be judged has false alarms.

10. A risk review device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 8.

11. A non-volatile computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

12. A computer program product comprising a computer program, or a non-volatile computer-readable storage medium carrying a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

Images